Your editor, git, and jump-to-definition all treat source files as the truth about your program, recovering meaning by re-reading them. That worked reasonably well when code was authored at human speed. It buckles when many AI agents write at once: names change on rename, locations change on insert, and content hashes change on every edit. This talk presents Fram, a fact-graph substrate where the database is the program, and Beagle, a typed Lisp that began life as a Racket #lang and can project source files from that graph. Delete a projected file, re-render it from facts, and it still compiles. I’ll show how stable identity turns edits into transactions, makes rename a fact update instead of a rewrite across reference sites, and changes the concurrency story for agentic programming, with wins, losses, and null results recorded side by side.

Racket and Rhombus support flexible idioms that pose challenges for type systems. Typed Racket, the gradually typed counterpart of Racket, uses occurrence typing to type-check programs whose control flow depends on run-time type tests. To alleviate the burden of annotation, Typed Racket also supports local type inference. Problems arise, however, when a Typed Racket program imports a macro from a Racket module that expands to complex code containing lambda expressions. Programmers cannot add annotations to generated parameters. Moreover, local type inference is not designed to infer types for lambda parameters. Therefore, the type system usually conservatively rejects the code. As a result, programmers often have to rewrite macros in Typed Racket. My ongoing prototype type inference system addresses the problem by combining occurrence typing and algebraic subtyping. In this talk, I will demonstrate how the new type inference handles patterns commonly seen in Racket programs.

Gathering with drinks and snacks.